Local models
A local model runs on your own computer, so replies cost nothing and your chats don’t leave it. It needs a capable graphics card beside Minecraft, and small models are less reliable at planning than the big hosted ones.
The AI connection screen already has two local presets, which need no key:
| Connection | Address it uses |
|---|---|
| Ollama (this computer) | http://localhost:11434/v1/chat/completions |
| LM Studio (this computer) | http://localhost:1234/v1/chat/completions |
Start the program, load a model, then type the model’s name into Model (these presets have no default model). Other local programs that offer the same kind of address can use Custom endpoint with an http://localhost address.
Coming: a guided local setup
Section titled “Coming: a guided local setup”Coming
A local route that works with as many local programs as possible, with a guided setup when you choose it. The plan:
- Find what you have. The mod looks for local programs on your computer without loading a model, lists what it found with its models, and explains how to switch a program’s server on if it is off.
- Presets for Ollama, LM Studio, llama.cpp, KoboldCpp and Jan, plus a custom address for the rest.
- Ollama done properly, through its own interface, so the context can be large enough for the mod’s prompts.
- Model suggestions by how much video memory your graphics card has, from MIT or Apache-2.0 licensed models only, after they pass the mod’s own tests.
- Download pages for the programs, opened through Minecraft’s own confirm-link screen.
- Test connection, the only step that sends a chat.
No date is set. See the Roadmap.